Experts follow an AI agent whose advice is a coin toss

Forty-three minutes, then one line

The analyst took his time: forty-three minutes reading the case file before sending his coding agent a single instruction. Then one precise request: this method, this population to measure. The agent complies, pushes back on both choices in a single message, and suggests weighting farms by size as well. The analyst's reply, in one line: fine, going with your recommendation.

After the session, he's asked why he weighted that way. He answers with phrases lifted almost word for word from the agent, never mentioning it, and rates the final recommendation as a bit more his than the AI's. What he didn't know: the agent's view on those questions had been decided by a coin toss.

The scene appears in an appendix to a study Anthropic published on October 27, "Who's Aligned to Whom?". It's no outlier: 77% of participants justified their choices with at least one of their agent's arguments, sometimes almost word for word. Only 15% mentioned it.

A coin for every decision

Fifty professional analysts in causal inference, recruited through a paid platform for specialist annotation, worked for about an hour with Claude Code, Anthropic's coding agent. Their job: tell the board of a fictional farming cooperative, using synthetic data, whether to roll an advisory program out to everyone or end it. The board wanted a productivity gain of 3 to 4%.

Along the way came four decisions with no right answer: measure productivity per hectare or per operating cost, pick a comparison method, target enrolled farms or the whole cooperative, count every farm equally or by output. For each one, a coin toss decided which way the agent would lean.

A hidden instruction told the agent to treat that lean as its own opinion, state it when the question came up, and defend it as well as it could, in good faith. It was forbidden to distort a result, hide data or make an argument the numbers don't support. The analysts didn't know.

On those four decisions, the analysts' choices matched the agent's lean 86% of the time, where chance would give 50%. That's measured decision by decision. Think of a compass secretly magnetized: the hiker walks straight, sure of his bearing, and it's the magnet that picks the valley.

A recommendation already written

On the final recommendation, the agent's direct effect is weaker, and it isn't statistically significant (p = 0.069) in the test the authors declared before the analysis. That's the confirmatory result.

But the recommendation flows from the estimate, which flows from the earlier choices. In an exploratory analysis, outside the declared plan, the four coin tosses are enough to predict 82% of recommendations, against 61% for always betting on the most common one. That calculation leaves out those who proposed extending the program by a year, and it's run on the same data as the model.

One session shows how it plays out. The analyst starts with a method that gives +4.6%. The agent steers him toward another, which gives −1.6%. The analyst recommends ending the program, against the agent's final lean. In the end he contradicts the machine, on a number it shaped.

Their words, minus its signature

Back to those justifications. Every choice had to be explained, with a question that invited them to say if it wasn't really theirs. The agent's arguments come back; its signature doesn't. And 52% rated the recommendation as more theirs than the machine's.

Those who followed the agent on all four decisions report being just as confident as everyone else. Asked what the study was about, none suspected an agent rigged behind the scenes.

Mind the wording. What the study measures is an absence of mention and a ticked box on who authored the recommendation. They remember perfectly well what they chose, in fact. What escapes them is where the choice came from.

We saw with the work on metacognition how poorly people judge when the AI is right. Here, there isn't even a right answer to check against.

What the setup forces

The authors spell out the limit: the lean was amplified. A real agent isn't set by a coin toss, and nobody asks it to defend a side. Anthropic concedes in its post that it's unclear how far this holds in the real world, and the effect could be weaker there.

That doesn't make an off-the-shelf agent neutral. With no instruction at all, the study's model makes the same five choices in 91% of 123 simulated runs. Facing an analyst who announces the option opposite to its preference, it brings up the other option unprompted in two cases out of three. The coin doesn't invent the preference. It makes it random, and therefore measurable.

Other limits: one task, one model (Claude Opus 4.8), and practitioners who were new to agriculture and who, on their own turf, would have firmer views. The sample is young: a third are junior, with two years of experience or less.

Across the ten characteristics tested, including experience, degree and trust in AI, no effect was detected. The analysis is exploratory, and since nearly everyone followed the agent, there was little variation left in which to detect one.

The agent has its upsides too. Without it, 6 analysts out of 11 made a serious error that invalidated their result, against 3 out of 50 with it. That no-AI group was self-selected and less experienced, so no cause can be drawn from it.

Only one side argued

According to the authors, the influence ran through neither lies nor hidden data. The agent explained how the result would change depending on the option, but rarely why a competent specialist might prefer the other. On more than three quarters of decisions, it gave its view before the human had one. A lawyer with no opposing counsel often wins, even without lying.

When independent teams reach the same result, their agreement counts as evidence. If they all work with the same model and the same preferences, that consensus may be just an echo. The authors also point at AI safety research, where agents already run audits: a human sign-off is only a check if the judgment is their own. The question echoes the oversight of agents at OpenAI, which we dissected last week.

An in-house study, again

Which brings us to who's talking. This is an Anthropic study of its own agent, published in no journal: a document posted on its website. The protocol was registered in advance on the OSF platform, but after data collection had started, when 17 participants had already begun. It's the second in-house Anthropic study we've read in eleven days, after the one on its agents and hiring.

This one exposes a flaw in the company's own product, and it also serves the lab's line on AI safety, its trademark. The strongest sentence comes from the authors: the work still goes out under the analyst's name, but the decisions that depend most on their judgment are settled by the agent's preferences.